+

WO2006107624A2 - Systeme et procede permettant d'extraire, de detecter, de differencier et de localiser une signature acoustique - Google Patents

Systeme et procede permettant d'extraire, de detecter, de differencier et de localiser une signature acoustique Download PDF

Info

Publication number
WO2006107624A2
WO2006107624A2 PCT/US2006/010892 US2006010892W WO2006107624A2 WO 2006107624 A2 WO2006107624 A2 WO 2006107624A2 US 2006010892 W US2006010892 W US 2006010892W WO 2006107624 A2 WO2006107624 A2 WO 2006107624A2
Authority
WO
WIPO (PCT)
Prior art keywords
signals
signal
signature
atoms
dictionary
Prior art date
Application number
PCT/US2006/010892
Other languages
English (en)
Other versions
WO2006107624A3 (fr
Inventor
Jeffrey M. Sieracki
Original Assignee
Sieracki Jeffrey M
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sieracki Jeffrey M filed Critical Sieracki Jeffrey M
Publication of WO2006107624A2 publication Critical patent/WO2006107624A2/fr
Publication of WO2006107624A3 publication Critical patent/WO2006107624A3/fr

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00

Definitions

  • the current invention relates to pattern and signal recognition and discrimination. It further relates to acoustic signal processing, and more particularly to sensing systems in which one or more transducers receive a sound, vibration, or similar signal that must be detected and identified in an automated fashion. It further relates to systems in which certain signals or classes of signals must be discriminated from one another. It still further relates to localization of the source of such signals.
  • This invention utilizes certain methods and systems previously disclosed in US Patent Application Serial No. 10/748,182, filed December 31, 2003 and entitled "Greedy Adaptive Signature Discrimination System and Method" (hereinafter referred to as the GAD Application).
  • Acoustic signals are derived from sensing the vibrations of a medium or object with an appropriate transducer. They may be recorded using microphones, hydrophones, or pressure sensors, or sensed indirectly through RF or laser vibrometry, optical interference, moving encoders, magnetic coils, or various other means. Acoustic signals typically comprise a one-dimensional time-varying signal that may be represented as a function, a voltage, a current, a pressure, a position, or so on, or may be sampled into a vector of digital numbers. Acoustic signals may also be grouped into multi-channel systems, which arise, for example, when multiple transducers are arranged to record the same signals using sensor arrays.
  • a "signature” is a pattern within a signal or data stream that can be associated with a condition of interest in the signal generating system.
  • Acoustic signals may be indicative of a state of affairs: A failing transmission, a knocking motor, an irregular heartbeat, congested lungs, rainy or windy weather, the strain on a suspension mechanism, and the proximity of talking people are a few examples.
  • Acoustic signals may also indicate the occurrence of an event: Breaking window glass, a roll of thunder, the report of a gunshot, the passing of footsteps, the failure of a bearing, etc.
  • Acoustic signals may also be used to identify their source: Mr.
  • the signal may be mapped to a class identity (e.g. is it a truck or a car?) or to a unique identity (e.g. is it Mr. Jones' car or Ms. Smith's car?).
  • Acoustic signals may also be utilized to identify the location of the source. Acoustic signals of interest will arise in many fields and numerous specific examples will be obvious to one skilled in a particular field.
  • the human auditory system is extraordinarily good at this sort of detection and discrimination. Yet creating a reliable automated equivalent remains a challenge.
  • a useable method must address several related goals: the signals must be translated into a representation that allows their manipulation and comparison; classes of signals must be compared in order to ascertain and extract characteristic signatures; a detector / classifier must be created to recognize signatures in a way that is robust to noise and environmental issues; and detected signatures must be localized in space.
  • the GAD Application discloses a suite of methods that can accomplish these goals when embodied in the proper context.
  • GAD refers collectively to the Greedy Adaptive Discrimination methods disclosed in the GAD Application.
  • GAD comprises several aspects, including a Simultaneous Sparse Approximation ("SSA") algorithm referred to herein as the “GAD decomposition algorithm” or “GAD SSA”, together with a system of signal representation and methods of processing that are reintroduced introduced as required in the text below.
  • SSA Simultaneous Sparse Approximation
  • GAD SSA Simultaneous Sparse Approximation
  • GAD Engine the operational elements of the GAD Application are embodied in a separable module, referred to as a "GAD Engine”, which can be utilized in various aspects to achieve signature processing.
  • “Sparse Approximation” is a term of art that refers to representing a potentially complex signal as the sum of a relatively small collection of component elements. "Simultaneous Sparse Approximation” is therefore the representation of each member of a group of signals in terms of a common, relatively small, collection of component elements. As disclosed in the GAD Application, the GAD decomposition algorithm in some embodiments permits the common collection of component elements to be similar rather than absolutely identical, thus increasing the utility of the idea.
  • SSA as used herein includes all variations of GAD SSA contemplated by the GAD Application as well as any similar or equivalent decomposition methods that may arise in the art.
  • the present invention defines certain embodiments of GAD that are applicable to acoustic signal analysis, along with certain refinements and additional complimentary methods that may be utilized in building deployable acoustic sensors and processors.
  • the methods and embodiments will also be useful in other applications in which similar needs arise.
  • the present invention comprises a novel method of processing, analyzing, and comparing acoustic signals in order to facilitate signature detection and localization.
  • the process begins with collecting acoustic data that is representative of the signals to be detected.
  • the data is normalized so that individual recordings are approximately comparable, and it is further divided into classes.
  • Each class comprises multiple recordings of a particular event of interest.
  • a simultaneous sparse approximation is performed on the data and parametric "mean" signals are generated for each signal class.
  • the method incidentally corrects for and removes parameter jitter between signals.
  • the parametric means comprise a collection of time-frequency atoms that represent a "typical" signal in the class.
  • the parametric means may, in some embodiments, be compared to each other in order to further reduce the dimensionality of the representations. For example, only those signal components that distinguish between classes may be kept and other components, common among the classes may be discarded. In certain embodiments, the components may be diagonalized in .order to achieve an orthogonal representation. In any case, by noting components that distinguish between signal classes, and/or noting class-typical values of components that are common among multiple signal classes, the invention establishes unique signature discrimination criteria.
  • the deployed sensor will utilize extracted parameters from the signatures to define a spectral filter corresponding to each signature.
  • the deployed sensor will directly utilize the collection of atoms that describes the signature, by comparing these to a similar analysis of any new signal.
  • One embodiment of such a detector is to generate a dictionary that contains compound atoms representative of the signatures of interest and utilize a nearest neighbor metric.
  • the parametric means will contain enough information to reconstruct an "average" signature signal in the original time domain. This reconstructed signature signal or the collections of signature components may be compared with any new signals by a host of well-known means, or by novel means elaborated in the GAD Application and below.
  • the problem of localization is addressed in two ways: by comparing amplitudes and by comparing time and phase information. All else being equal, a louder signature is generally closer than a quieter version of the same signature. R- squared power relationships are utilized to make rough comparisons, and these can be normalized to the environment. Furthermore, when multiple transducers with known, separate positions record the same signature signal, timing and phase information may be utilized to predict the source location.
  • the invention makes available time and amplitude information that may be utilized by other known methods, or utilized by specific processes disclosed here.
  • One embodiment of the present invention works by sequential operation.
  • a signal recorded by one or more sensors is normalized and compared to the signature database. If there are multiple transducers, several possible embodiments are available, hi a first, each channel is compared individually to the database and a weighted decision metric makes the final determination.
  • the signals are cross-correlated to phase align them and the summed (or averaged) signal is compared to the database.
  • the signals are analyzed using the GAD sparse approximator (as disclosed in the GAD Application) and phase-aligned and de-jittered by taking a "parametric mean”; the "parametric mean” signal is then correlated to the dictionary.
  • the last method offers specific advantages, for example, when distributed sensors are only approximately located or have free running data clocks, both of which introduce unknown variation into timing and position information.
  • a signature is determined to be present and (if necessary) has been properly classified, it is then located within the recordings from each individual channel.
  • the relative phase, timing, and energy (amplitude) information is analyzed across channels to localize the source.
  • the signal may be located within each channel by a host of well-known methods, including for example cross-correlation or pattern search. It may also be located, in certain embodiments, by extracting parameters directly from the GAD sparse approximator output rather than performing an additional calculation.
  • location characteristics may be learned by the system in the same manner as signatures are discovered by exposing the sensor(s) to a variety of different signatures in different locations.
  • Fig. 1 illustrates an elementary acoustic sensing system in an ideal environment, and a multi-channel acoustic sensing system in a complex environment.
  • Fig.2 is a schematic representation of a GAD engine.
  • Fig.3 is a flow diagram of a first method of signature extraction.
  • Fig. 4 is a flow diagram of an improved method of signature extraction.
  • Fig. 5 is a flow diagram of a general detector / classifier mechanism, together with a flow diagram for a method of implementing feature detection with the subject invention.
  • Fig. 6 is a flow diagram of a two stage detector / classifier, utilizing separate processes for discovering signal component atoms and signature combinations.
  • Fig.7 is a flow diagram of a simultaneous, multi-channel detector / classifier.
  • Fig. 8 is a flow diagram for an adaptive detector / classifier that matches signals against prior samples rather than to an extracted signature set.
  • Fig.9 is a flow diagram for recovering time localization information from a multichannel SSA analysis.
  • Fig. 10 is a flow chart showing a process for training a subject system to localize signal sources by example.
  • Fig. IA depicts a simple, idealized schematic of an acoustic sensing system.
  • a sound source 10 radiates acoustic energy through a medium and a microphone (or other appropriate transducer) 11 converts the detected vibrations into electrical energy.
  • a processing device 12 amplifies, preconditions, and evaluates and/or stores the resulting signal.
  • Fig. IB depicts a more realistic scenario and illustrates the complexities of the real world problem.
  • a source 10 radiates acoustic energy.
  • other media with varied acoustic properties now interrupt the smooth transmission medium. Examples include inhomogeneous regions such as 13 that absorb and refract energy, possibly in frequency dependent ways; and reflective objects 14 that bounce and scatter energy.
  • the multi-path wavefronts combine at the microphone (or other transducer) 11 to obscure the source acoustics.
  • This second picture depicts optional additional transducers that independently sample the acoustic energy at multiple points.
  • the transducers 11 may be identical, or may differ from one another in operation or response characteristics.
  • the transducers may also be randomly spaced, in which case 12 must know their relative locations in order to extract coherent information through comparison.
  • the transducers may even be in motion, relative to each other and/or relative to the acoustic source.
  • one mechanism for aligning the signals temporally and calibrating the system is to transmit localizing information (such as differential GPS) to 12.
  • Another method is to utilize a known source of acoustic emissions and extract phase and time differences from the signal arrival times at each transducer.
  • Another novel method afforded by the methods disclosed in the GAD Application and below, is to extract a matched signature pattern from uncontrolled emissions using the subject GAD methods and extract the timing information from the parameters of matched atoms. (See below.)
  • the dotted region 26 represents a "GAD engine”, comprising a Simultaneous Sparse Approximator ("SSA") 22, a structure book memory system 23, and one or more discrimination functions 24 that operate on the structure books.
  • the SSA 22 takes as input a plurality of signals 21 and produces a structure book for each signal.
  • the output of the SSA will comprise one or more structure books sub-selected or otherwise processed as described hereafter and/or previously in the GAD Application.
  • a structure book describes a linear decomposition of the signal and comprises a list of coefficients and a corresponding list of atoms for the decomposition.
  • One way to write the structure book is as a set of ordered pairs, (ai t gi(t) ); however, we need not record the atom gj(t) itself.
  • Descriptive information stored in the structure book may comprise the atom itself, a coded reference to the atom, or one or more parametersthat uniquely define the atom; memory efficiency, speed, and convenience of accessing the atom and/or its properties will dictate the choice.
  • the gj(t) belong to a dictionary D of prototype signal elements, more fully described in the GAD Application.
  • the dictionary is an intrinsic element of the SSA as described in the GAD Application and hence is not shown explicitly in all figures herein.
  • the dictionary may be implicit rather than an explicit separable component. Where a dictionary becomes relevant to the functional description of a figure, it is broken out explicitly from the SSA block for explanation purposes.
  • structure books are created relative to a dictionary D and subsequent operations are performed based on this implicit relationship.
  • a structure book may be recast into another representation by mathematical projection operations well known to those skilled in the art, in which case the elements gi(t) and the coefficients aj used in the structure book may change. In some cases these new gi(t) will belong to the original dictionary D, in other cases a new dictionary will be used.
  • the illustration or lack of illustration of a particular dictionary D in describing a subject process herein is for convenience and is not a limitation of the invention.
  • the SSA produces structure books for each signal in the input collection 21 such that the atoms of any structure book may be compared directly to those of any other.
  • the atoms may be identical for all signals in the collection.
  • GAD SSA is also able to produce atoms that are "similar" as judged by the algorithm rather than identical; this novel feature is advantageous in many implementations because it allows the algorithm to automatically account for noise, jitter, and measurement error between the signals 21. See the GAD Application for details on methods for implementing SSA. Algorithms that produce similar simultaneous approximations for a group of signals may be substituted with appropriate adjustments.
  • the atoms selected will vary depending upon the SSA implementation.
  • the output of any such SSA may be further processed (e.g., to orthogonalize the atoms in the structure books) without departing from the spirit and scope of the present invention.
  • the general GAD SSA disclosed in the GAD Application permits the range of "similarity" between atoms across structure books to be controlled by setting a search window for each of the parameters of the dictionary.
  • the windows may be fixed in advance for each parameter, or may be adapted dynamically.
  • One adaptation that is sensible, for example, is to adjust the search window according the classical uncertainty principal; for example, appropriate search windows (and step sizes) for time and frequency might co-adjusted based on the time or frequency spread of the atom.
  • the purpose of the variation is to associate similar by not identical atoms in an automatic fashion, and many possible windowing schemes will fall within the general mechanism disclosed.
  • a Gabor dictionary as defined in the GAD Application, comprising modulated, translated, and scaled Gaussians, combined with Fourier and Dirac delta bases.
  • This example dictionary in no way limits the scope of the invention, and any other reasonable collection of prototype signals may be substituted, including in certain embodiments a dictionary of random prototype signals.
  • the dictionary may be orthogonal, such as a Fourier basis, or not. It may be redundant, such as a collection of wavelet packet bases. It may also be highly redundant, as is the Gabor dictionary described.
  • the post SSA discrimination functions depicted generically as block 24, are a class of operations that utilize one or more existing structures books in order to create new structure books. Certain operations will expand or reduce the number of structure books recorded in memory 23, while others will simply replace or modify existing ones.
  • One such basic operation is summation, in which the coefficients of one or more structure books are added together to create a new structure book. This idea can be used to generate an "average” by summing a set of N structure books and dividing summed coefficients by N.
  • the parameters of each atom in the various structure books may be averaged as well to produce a "parametric mean” structure book. See the GAD Application for more details.
  • Two structure books can likewise be subtracted in order to calculate the difference between them.
  • Another example operation is thresholding.
  • the coefficients of a structure book are compared to a set value, and the atom (or in certain embodiments the entire structure book) is discarded if it is below (or above) that value. Thresholding may also be based on parameters of the atom; it may likewise be based upon absolute values of coefficients or parameters; or may be based upon relative variance from a defined value or distance from a group mean. This latter may be extended to cluster analysis techniques. Again, see the GAD Application details for other details; specific examples are utilized below.
  • Various sorts of thesholding may be used to either reduce the dimensionality of the representation by eliminating atoms, or it may be used to classify signals based upon the different values of their coefficients or atom parameters. Note that using absolute (or squared or other even powered) measures rather than signed values is important when magnitude or energy of a component is established as a decision criteria, whereas using actual valued (or odd powered) measures retains sign of the comparison and is important when the direction of excursion is established as a decision value.
  • structure books may be treated by other methods to reduce dimensionality. For example, if the structure books are drawn from a redundant dictionary, it may sometimes be useful to orthogonalize the representations by generating new atoms using a gram-schmidtt type process. It is also possible to operate on a structure book to collapse one or more dimensions of multi-parameter atoms. This is illustrated in the GAD Application for the case of collapsing a class of signals across time in order to compare certain classes of time-correlated signals with randomly correlated background activity.
  • the GAD Engine components define a very flexible tool for manipulating and discriminating signals. Further examples are illustrated in the GAD Application and below.
  • One unique benefit of the certain SSA methods disclosed in the GAD Application is that the resulting structure books may be averaged, subtracted, or otherwise manipulated and any derived structure book retains sufficient information to reconstruct a representative, model signal in the original signal space. In particular, this means it is possible to calculate a parametric mean of a class of signals and then reconstruct a "typical" signature signal from that data for further analysis, comparison, etc.
  • Acoustic data collected by means of a transducer system must in general be preconditioned.
  • the electrical data will typically be amplified to make it usable in further processing stages. It will also be limited, in order to prevent damage to the system from excessive input levels. It may be compressed with an automatic gain control system, or other means in order to normalize the energy envelopes. It may in addition be filtered to remove noise, or to highlight frequency bands of particular interest in the application. It may also be gated in various well-known ways to remove noise.
  • These and other preconditioning operations may be accomplished either in analog circuitry or digitally after sampling the signal. Many systems will utilize both mechanisms, because on the one hand it will in general be necessary to optimize the analog signal for sampling while on the other, digital operations are often more flexible and convenient.
  • sampling must be at a sufficient rate to exceed the Nyquist requirements of the highest frequency components of interest in the data.
  • Bit depth must be sufficient to allow dynamic range for likely excursions of the acoustic data, and for discriminating signatures in a noisy environment.
  • Multi -rate and multi-bit-depth systems may be employed in order to economize or speed operation in certain embodiments; for example, digital filtering, compression and limiting may be applied before down-sampling the signal into a more compact data type for comparison processing and other operations.
  • Preprocessing is not a strict requirement of the subject invention, and its necessity and details depend upon application. The different aspects herein discussed will be integrated only as appropriate to a particular embodiment.
  • acoustic signatures of interest would be well known, easily recognizable patterns, which are repeated nearly perfectly each time they are emitted. In rare instances, this may hold. However, in practice, even the highest tolerance mechanically produced sounds and vibrations will vary from instance to instance. Diffusion and intervening reflective and refractive media interfere with and dynamically distort the signal; background noise arises; measurement error and time jitter occurs; and so on. Furthermore, it may not be at all clear a priori which characteristics of a signal distinguish it from either the background noise or from another signal of interest. Therefore, signatures must be identified and extracted from collections of data.
  • Fig.3 One subject process for signature discovery is illustrated in Fig.3. It begins with sampling a collection of audio segments, each of which is preconditioned to contain one similarly sized instance of the sound of interest. The segments may be further preconditioned as described in the previous section.
  • the audio segments are divided into classes, where each such class is a collection of sample segments of only one type. For example, if we were training the system to distinguish 6-cylinder from 8-cylinder engines, then one class would contain only 6-cylinder engine samples and the other class only from the 8-cylinder samples. Likewise, if we were training the system to distinguish a hand clap from a gunshot, then one class would contain samples of handclaps, and the other samples of gunshots.
  • each individual class is analyzed separately by performing the following operations: First the SSA 32 creates a collection of matched structure books (33 or 34) from the signals (30 or 31). Next a parametric mean operation 35 is performed on the matched structure books to produce a signature structure book for each signal set (36 and 37). If the SSA produces structure books with identical atoms for each signal in a class, then the parametric mean reduces to just an average of the coefficients. However, if the general GAD SSA is applied to produced similar rather than identical structure books for the signals within a class, then the parametric mean will include an appropriate averaging of the parameters of the respective atoms.
  • the structure books 36 and 37 may be compared with each other at this stage to reduce the dimensionality of the representations by extracting discriminating elements. This is illustrated with the optional dotted box 38.
  • the idea is better facilitated by a second signature extraction method described in Fig 4; hence, the sub-process of block 38 is addressed below.
  • each coefficient ajand each atom gi(t) are directly comparable for a fixed i only within the class the structure book is associated with; but atoms in 36 are not in general the same as atoms in 37. (There are special cases, e.g. use of orthogonal dictionaries, wherein 36 and 37 will be atom-compatible.)
  • the resulting parametric mean structure books 36 and 37 produced by the extraction process may be utilized to represent the signatures in a compact form.
  • the structure books 36 and 37 may be used to reconstruct model signals for each class.
  • the process of reconstruction, block 39 involves summing the components represented by the structure books, in the manner of the linear combination described in the GAD summary above and further in the GAD Application.
  • the output of the process is a "signal" representative of each class. If reconstructed, then these "typical" signals (40 and 41) may be construed as signatures and lie mathematically in the same signal space as the original acoustic; hence, they may be directly manipulated just like an acoustic signal.
  • the reconstruction is illustrated using dotted boxes, since it is an optional part of the process. However, because the process just described generates structure books 36 and 37 that are not directly comparable, the optional reconstruction of 46 and 47 will often prove useful in further processing. Use of these signatures for detection is addressed below.
  • FIG 4 A second process for signature discovery is illustrated in Fig 4. It also is illustrated for only two classes, but may be similarly extended to any number and this schematic should not be interpreted to limit the invention.
  • the process begins similarly to that above, by sampling and precondition audio segments. Again, the segments within the collection are divided into classes based on their origin.
  • the method illustrated in Fig 4 differs, however, from the previous process in that all classes are combined and analyzed simultaneously by the SSA 32.
  • the resulting analysis has the important advantage that the structure books (42 and 43) for both (all) signal classes are now directly comparable. Each coefficient ai and each •atom gi(t) are directly comparable for a fixed i, regardless of which class the structure book is associated with. This novel advance was introduced in the GAD Application.
  • Optional block 38 again illustrates that the structure books 44 and 45 may be compared and further analyzed in order to reduce dimensionality or otherwise improve their usefulness in later detection and discrimination applications.
  • Signature structure books generated by either the first (36, 37) or second (44,45) method may be compared; hence, the following applies to structures books generated by either of the illustrated processes, or by any similar process.
  • the process diagramed in Fig. 4 guarantees that the structure books 44 and 45 share atoms, and hence simplifies and speeds the inter-comparison.
  • an additional process must be applied within 38 to match "similar" atoms with one another across the two sets. This may be archived by nearest neighbor matching between the sets.
  • Another option for relating one independent atom to another is to transform the signals into the time-frequency plane (or other higher dimensional representation) and compare data there.
  • Still another option for relating atoms to one another is to drop one or more parameters; for example, time and phase information may be dropped to form larger equivalence classes if this information is not germane to the signals of interest.
  • atoms that are matched to other similar atom in other structure books will be eliminated from both structure books, since they do not contribute to the discrimination.
  • similar we define "similar” as those SSA matched atoms that have similar energy (coefficient magnitudes) in both structure books. Atoms that appear strongly in one signature structure book and weakly in the other are retained. This may be achieved by various search methods and is related to a set theoretic subtraction (or "relative complement") operation, fl ⁇ f2, where fl and f2 are the signature structure books of interest. Another embodiment of this idea is to take a difference between the matched structure books of the two signatures, subtracting the energies in one from the energies in the other.
  • a second embodiment is to square each coefficient before the subtraction in order to ignore the sign of the atom. This should be decided with caution, since in certain cases one signal class may differ from another signal class simply in the sign of one component. (Consider, for example, an embodiment that analyzes phase-encoded signals such as used in BPSK or QPSK.)
  • the appropriate difference fl ⁇ f2, fl - f2, fl 2 - f2 f 2 for ie [O..n], etc., ' between signatures, however arrived at, will represent those components that strongly present in fl and not in f2.
  • the reverse difference, e.g., G - fl represents those components that are strongly present in H by not fl. Additional details appear in the GAD Application.
  • One embodiment for distinguishing one signature from a large collection of signatures will proceed similarly as follows: A difference is formed between the first signature fl and each other signature in the set, as described above. The resulting difference structure books are then analyzed to establish which, if any, atoms they have in common. In certain embodiments, a classification decision tree can be created from the structure books and a detector may be designed based upon the result. (See below.) In another embodiment, each of the pair-wise reduced dimensionality signatures will be simply utilized independently in the dictionary of the detector. In still another embodiment, the set of pair-wise difference structure books will be merged. This last method is undertaken by taking the logical union of atoms of significant energy in the collection of difference structure books. This union represents the minimal set of component elements that must be detected and measured in a new signal in order to distinguish it uniquely from its peers. The process successfully reduces the dimensionality of the detection problem.
  • the explicit subtraction operation can be replaced with a sequential "search and compare" operation over the set of atoms and the set of signature structure books.
  • An additional method for reducing the dimensionality of the representation is to collapse atoms across one or more parameters.
  • a Gabor dictionary of acoustic prototype elements is naturally parameterized by time, frequency, scale, and phase.
  • phase of a component is unimportant, hence those atoms that are selected by the SSA may be treated as identical for analysis purposes if all other parameters are similar except for phase.
  • time of onset is unimportant; this may occur for example in continuous, stationary signature signals. Therefore atoms may be collapsed across time.
  • equivalency sets may be established, expanded, and applied in implementing the methods herein disclosed.
  • the structure books may be orthogonalized by a Gram-Schmidtt or similar process in order to express data in a minimal orthonormal basis. In certain circumstances, this will have benefits for detection, comparison, or other processing.
  • the structure books may also be transformed into a known representation, such as a Fourier or Wavelet basis. Orthogonalizing the representations may, for example, ease the removal of common components from two or more structure books.
  • the SSA methods disclosed in the GAD Application are easily specified to automatically generate orthogonal representations by utilizing an orthogonal (rather than a general) matching pursuit.
  • the structure books will "automatically" orthogonal. This is contemplated within the scope of the GAD Application. However, it is not always desirable. Redundancy often achieves more meaningful representation of the data, more options in later processing, less susceptibility to round-off and truncation errors, and can generate more robust representations in the face of noise and measurement error. Furthermore, orthogonalizaton can be slow; hence, if it is necessary it will often be efficient to wait until the dimensionality of the data has already been reduced. Orthogonalization may be achieved by using an orthogonal dictionary, by using an orthogonalizing SSA or by post-processing the structure book(s). All of these are contemplated.
  • optional process block 39 may be applied to reconstruct the structure books into signature signals, 46 and 47 by summing the represented structure book components. Whether or not process 38 has been applied to reduce or modify the structure books 44 and 45, reconstruction 39 will still operate similarly.
  • the methods described so far for signature discovery and training do not required that the signal be reconstructed.
  • the problem is instead being attacked in a low dimensional representation defined by the structure book(s). In many embodiments, this gives advantages of tractability and robustness.
  • the signature signals may be reconstructed from the structure book representations.
  • These time domain signals can be utilized directly by certain detector designs. (See below.) Time domain representations will be most useful in representing one-shot time course signals, or transients, where the phase and time become a component of the signature. When signals of interest are continuous and stationary, then dictionaries that disregard time envelope information will generally be more useful; and in some cases, disregarding phase will also be appropriate.
  • Background noise is a common and ever present problem in acoustic signal detection. Self-noise from microphones, amplifiers, and sampling circuits forms a relatively stationary backdrop for all recordings. If sensing and sampling equipment for training differs from that used for deployed detectors, compensation must be built into the system. Environmental background noise is an even more significant problem, since it will change from environment to environment, and from time to time within environments. Procedures may be followed to readjust detectors accordingly at training (signature discovery), on deployment (signature detection), and even dynamically over time. The procedures are related, and in many cases similar.
  • a single channel noise gate may be triggering by sound envelope patterns or thresholds; multi-band gates may also be employed. More advanced noise reduction may be achieved digitally by utilizing Fourier, wavelet, or other basis or frame decompositions and thresholding the resulting coefficients. Any of these may be utilized to pre-condition the signal during training and/or during detection, and will be applied within certain embodiments of the present invention.
  • An additional technique may be applied within the GAD engine.
  • a signature has been established for a class of signals and reduced to a parametric- mean structure book. (Or, if only one-signal has been sampled, simply suppose we have generated an appropriate structure book.)
  • energy e.g., squared or absolute values
  • RMS radio frequency fingerprinting
  • the "typical" background energy (or amplitude) of each atom may be utilized to threshold the energy (or amplitude) associated with atoms in any structure book. Those atoms represented with sufficiently large coefficients are considered to be above the level of background noise, and contribute to the signal of interest. Those that are weakly represented are removed, reducing the dimensionality of the structure book.
  • This process may be applied during training to eliminate recording system self-noise from derived signatures. It may be applied in any specific new environment to threshold signals and/or signatures against background noise. It may be applied periodically by user input, or by an automatic process, to continually recalibrate the system against changing background noise.
  • the detector When background noise increases in the field, the detector may be recalibrated to ignore atoms with energy below the background threshold. Any structure book representation of a signature may be thresholded and those atoms that are too small to detect may be eliminated for comparison purposes. This increases decision robustness and decreases processing overhead until the background noise again drops. The same process may be applied, either independently or simultaneously, to a newly sampled test signal.
  • atoms below noise threshold may be mapped to appropriate detectors by referring to their descriptive parameters - for example, a Gabor atom may be mapped directly to a filter of specific frequency and bandwidth.
  • Signal Detection, Discrimination, and Classification Detector / Discriminator mechanisms operate by examining the signal for features of interest.
  • Fig 5A illustrates the concept.
  • the new sample 50 is analyzed either simultaneously or sequentially by a set of feature detectors 52, 53, 54, the output of which is either a binary true/false, or a weighted value proportional to likelihood or magnitude of the feature in the signal. Any number of detectors may be included, only three are shown for illustration.
  • the outputs ar& compared 55 to determine if a signature is present, and/or classify it.
  • the feature detectors in various embodiments respond to frequency components of various bandwidths, scale components, phase components, time localized components, etc., in any combination.
  • a filter bank can be utilized to detect the presence of a continuous signal with certain characteristics.
  • features can comprise any characteristics that differentiate between atoms of the dictionary. Thus complicated features may be discovered and extracted by utilizing an appropriately constructed dictionary.
  • FIG. 5B illustrates further applicability of the present methods.
  • a signal 50 is matched by some process 56 to a dictionary 57 of possible choices.
  • This matching process may be an SSA, a matching pursuit, or another decomposition; it may be a correlation operation; or it may be any of numerous other methods known to those skilled in the art.
  • the process utilizes a dictionary of features, e.g. signatures or atoms of interest.
  • the output is structure book 58 (or equivalent), from which proportional matches to each feature (52,53,54, etc.) are extracted.
  • One method of detection and discrimination is as follows: Construct a dictionary of reconstructed signature waveforms. Find the cross-correlation of each dictionary waveform with each new sample window. Examine the peak value of the cross-correlation function and apply a threshold to determine if a signature is present. The best-correlated dictionary element is the likely candidate. The cross- correlation peak, incidentally, also provides arrival time information that may be utilized for localization. (See next section.)
  • the threshold(s) may be fixed, or may be adoptive to background noise as descried above.
  • Another method of detection and discrimination is as follows: Construct a dictionary of reconstructed signature waveforms, together with time shifts of said waveforms. Using a greedy algorithm, such as a matching pursuit, decompose each new sample signal into a sum of dictionary atoms. Threshold the resulting coefficients in order to determine if a signature is present.
  • a greedy algorithm such as a matching pursuit
  • the time-shift parameter of the signature provides arrival time that may be utilized for localization.
  • the threshold may be fixed, or may be adoptive to background noise as descried above.
  • the number of time shifts of each signature signal within the dictionary will be determined by the uncertainty principal, with (optional) further refinement in time occurring once a signature is approximately located.
  • This algorithm may also be coded by simply listing each signature once in the dictionary and taking time shifts as needed before each inner product comparison.
  • detection may proceed on one or more channels individually, and the results can be pooled to determine statistically the likelihood of positive detection.
  • the signals are combined with the following method: The signals from individual channels are cross-correlated to time- and phase-align them, by matching correlation peeks. The time-phase aligned signals are then summed (or averaged) and the resulting signal is compared to the database as in the same manner as any other one-channel signal. This idea may also be extended using the subject GAD engine to analyze multiple signal channels at once.
  • a method of simultaneous detection and discrimination is as follows: Construct a dictionary of reconstructed signature waveforms, together with time shifts of said waveforms. Using a SSA, decompose the signals simultaneously into matched structure books. By using the GAD SSA, time and phase similarity windows may be defined wide enough to accommodate different arrival time of the signature at different microphone. Threshold the resulting coefficients in order to determine if one or more signatures are present. Again, multiple signatures may be detected in the same window. The extra information available from simultaneously treating multiple transducers will, in certain circumstances, increase the robustness of the detection. Again, the time-shift parameter of the signature provides arrival time that may be utilized for localization. Again, the threshold may be fixed, or may be adoptive to background noise.
  • FIG. 7 A plurality of transducer channels 70,71,72 are analyzed simultaneously by an SSA 76, utilizing a signature dictionary 77.
  • the resulting matched structure books 78 are processed by taking a parametric mean 79, and the mean structure book is fed to a detector / classifier as elsewhere described (e.g., Fig 5 above, or Fig 6 below.)
  • the dotted boxes 73,74,75 indicate optional time delays that may be inserted to compensate for variations in arrival time at transducers so that the time-similarity window may be decreased in the GAD SSA to reduce false associations.
  • Such pre delays may be utilized as well in locating the sound source; this is addressed below.
  • the delays may comprise actual time shifts introduced before the SSA or may be coded as time shifts that occur on the fly during the comparison operations of the SSA.
  • Each method may be modified to utilize the reduced dimensionality structure books discussed above.
  • two signatures fl and f2 have been identified using the parametric mean process.
  • One embodiment utilizes a dictionary that includes fl ⁇ C, f2 ⁇ fl, and the intersection, fl n f2.
  • the signatures fl and f2 are represented in their entirety using separable components. Detection of fl ⁇ f2 and fl n£ together in correct proportions positively identifies the presence of fl in the sample. The detection is robust since all components of the signal are factored in, while the discrimination decision between fl and f2 is simplified since the common components may be ignored without taking additional steps.
  • Each method may be modified to utilize other groups of atoms rather than the complete reconstruction to represent signatures in the dictionary.
  • One example is to simply include each atom from each extracted signature's structure book in the dictionary. This embodiment will offer advantages in that the dictionary is compromised of standard atoms that, in certain circumstances, will lead to efficiencies of implementation. It may also reduce storage requirements since the dictionary can take advantage of redundancies. However, detected atoms must be mapped to the signatures. If we proceed by decomposing a new signal in the dictionary of signature sub components, we must then determine if the resulting decomposition is sufficiently close to one of the known signatures to qualify as a match.
  • Fig. 6 illustrates one embodiment of the process.
  • the new signal 50 is fed to a matching engine 61 (in this case a Matching Pursuit operation, but variations may be substituted) that utilizes a dictionary of atoms 62 as just discussed.
  • the resulting structure book 63 is matched with a nearest neighbor metric 64 against a signature dictionary 65.
  • the signature dictionary comprises the combinations of atoms and their relationships (e.g., signature structure books) that are within the comparison set.
  • the nearest neighbor search may be performed by numerous methods, well known in computer science, with the goal of finding a signature structure book in the dictionary 65 that minimizes some metric. Typically this metric quantifies a "distance" between the test structure book 63 and any given signature structure book in the dictionary 65.
  • a vector norm on the difference in coefficients of matched atoms is one functional embodiment.
  • a vector norm will be applied to the atom parameters to quantify how close one atom is to another; thus dissimilar atoms can be grouped approximately for comparison purposes.
  • the illustrated process next makes a decision as to whether the match was close enough 66 to reliably suggest a good fit. If yes then we report detection along with the matched signature, otherwise we report no match. In various embodiments, this test 66 will be performed by thresholding, by statistical inference, and by other heuristic means.
  • the atom dictionary in some embodiments will be the entire analysis dictionary; this has the advantage of analyzing each new sample in the same manner as the training corpus, and allowing noise in the signal to be mapped away from the atoms need to detect signatures.
  • the reduced dictionary containing only those atoms necessary for signature detection will be include in the dictionary, speeding processing and simplifying the nearest neighbor comparisons.
  • the GAD SSA is a greedy algorithm, selecting best fits first; a compromise therefore is to include all atoms selected in the process until it was terminated, whether or not those dictionary atoms are associated with a particular signature. This ensures that larger noise components are mapped away from the signature sets.
  • Fig. 8 A collection of signature samples 30 of one or more classes is analyzed together with a novel sample 50.
  • the SSA engine 81 (such as that described in the GAD Application) produces a collection of matched structure books 82. These structure books are then analyzed by means of comparing the values associated with atoms of the novel signal's structure book to those of the remaining structure books. If they are within a limit defined by the variance range of the known matches, we declare a match, otherwise no match.
  • the variance analysis may be a standard statistical test (e.g. a Z-score) or any of a variety of measures that compare one value to a group of others. It may be as simple as a threshold to test if it is within the spanned range of variance.
  • the method may include only a single class of signatures in 30, in which case the test 83 is binary. It may also include a plurality of signature classes, in which case the test 83 is modified to perform a nearest neighbor on the segregated structure books in the manner described for Fig 6.
  • This novel method has certain advantages over the other systems. In particular, no training is required and no dictionary of abstracted signatures is required. Among other benefits, this also allows a deployed system the ability to learn by exposure and accommodate the environment by example. In this mode, as each new signal is classified a match, it may be added to the corpus. The method also allows us to create a new "classification" engine building block that can be used without pre-training in larger processing systems.
  • amplitude and phase There are two primary means of determining the location of an acoustic energy source from the sound-pressure on one or more transducers: amplitude and phase.
  • a directional microphone for example, one may scan an acoustic field to find a peak in amplitude response.
  • two or more microphones one can compare volumes to predict direction; the prediction may give a sufficient result, or may guide a scan that better localizes the source.
  • the difference in phase (or more generally, relative time delay) of signal arrival between two or more microphones also predicts direction.
  • the microphones may be steered to localize the sound, or a best estimate may simply be extracted from the time arrival times without moving the microphones.
  • the invention utilizes novel SSA methods to extract this information from noisy signals, thus enabling localization.
  • the directional sensitivity of an array of transducers may be electronically steered by adjusting relative time delays before combining their signals.
  • This well- known technique is termed beam forming. It may be used to steer and/or focus transmitting or receiving sensor elements.
  • many devices of prior art perform a virtual scan of the space of interest. In effect, they adjust relative time-delays in a controlled fashion to study local portions of the sound field, while searching for a peak in the signal of interest.
  • the systems and methods disclosed in the GAD Application and in the present document afford unique benefits in aiding this sort of search.
  • a scanned beam- forming array is fed into a signature detector, as described in previous sections. Positive detection then identifies not only the signature of interest, but also its relative location (or direction.)
  • a second combination of techniques is as follows: One or more individual transducers of the array are analyzed for signatures. On positive detection, the recognizable components of the signature are extracted. The beam-forming array is then scanned across the sound field looking for the source of the extracted components. Embodiments of this technique often have faster update cycles than those of the preceding technique, because the scanning process is not initiated until a signature is identified in the first place.
  • One advantage of the subject invention and the detector designs described is that they can extract a precise model of the signature as it actually appears in the environment, and even be utilized to reconstruct the waveform of interest exactly.
  • the "scanning" process may be real in certain embodiments, or may comprise a virtual scan of the multichannel data recorded form the transducer arrays.
  • the virtual scan of stored data has two advantages: (1) The target signature is known precisely since it was extracted from the same data, and (2) if the target signature is a transient or non- stationary signal it can still be recovered from the saved data.
  • Source localization procedures are not limited to pre-constructed phased arrays of transducers. Any set of transducers with known relative locations can be utilized to discover location information. The relative delay methods may be applied if the spacing exceeds the wavelengths of interest, however and the general term "time” rather “phase” delay applies.
  • Arrays of sensors need not be fixed. Ad hoc arrays may be constructed from available data sources; transducers may be placed as needed or scattered by some means. If the transducers are not at fixed locations, one mechanism for aligning the signals temporally and calibrating the system is to transmit localizing information (such as differential GPS, or inter- sensor acoustic pulses). Another method is to utilize a known source of acoustic emissions and extract phase and time differences from the signal arrival times at each transducer. Amplitude information may likewise be extracted, and the arrays may be calibrated (or cross checked) by applying R-squared dispersion models.
  • a further method of discovering transducer locations enabled by the methods of the GAD Application is to utilize GAD to analyze an unknown source at a known location. Once a signature is extracted by processes discussed above, one can then look for the same signature across channels an extract time/phase information (and/or amplitude) to calculate position information for the sensors.
  • a third combination of techniques afforded by the invention is as follow:
  • One or more individual transducers of the array are analyzed for signatures.
  • the recognizable components of the incoming signature are extracted to a structure book.
  • An approximation of the detected signature is reconstructed.
  • Each channel is then examined to extract time of arrival of the signature elements. If the signature cannot be reliably identified in a given channel then that channel is dropped from consideration. Relative arrivals times (and/or amplitudes) are then compared among successful channel(s) in order to insure consistency; again, problem channels may be dropped from the pool.
  • timing information from remaining channels is combined to extract an estimation of source location.
  • Extracting relative arrival times is straightforward when the signal of interest is loud in comparison to the background.
  • the reconstructed signature can be compared to each channel by cross-correlation, and the resulting peaks will occur in each channel at the relative time delays.
  • the invention affords another novel variation of this idea: rather than reconstructing the signature in its entirety, one or more component elements of the signature structure book may be analyzed separately. This process improves over the prior art, in that it automatically compensates for possible frequency dependant time or phase delays introduced by the environment. By allowing the components of the signal to vary in time relative to one another, and then averaging the resultant estimates of arrival time, it is possible to achieve lower channel rejection rates while minimizing overall error.
  • a greedy algorithm for combining channel information proceeds as follows: Select the channel with the best signature candidate in the data and extract an arrival time estimate, select the next best channel and extract an arrival time estimate, etc. As the number of channels exceeds the required number for localization, compare for consistency, dropping channels that are inconsistent with the majority of the data points.
  • Figs. 9A and 9B An additional method of directly extracting time information is also afforded by the methods of the GAD Application in the presence of multi-channel data. Signature detection and discrimination may proceed using a simultaneous GAD process on all (or a subset) of the transducer signal segments at once. This procedure is illustrated in Figs. 9A and 9B.
  • Fig. 9A illustrates the basic idea: Signals from a plurality of transducer channels 91, 92, 93 are analyzed simultaneously by and SSA 94, utilizing a signature dictionary 95. The resulting matched structure books 96 are processed by comparing phase and time information between matched components in 97. The differences across channels can be averaged, thus extracting an estimate of the relative delay between each channel. From the time delays, source position information can be calculated in a well-understood manner.
  • the time- similarity range of the GAD SSA may be set to accommodate the worst reasonable delay.
  • a time delay may be introduced in each channel, in order to allow the system to be "scanned" in the manner of a phased array: that is, sensible time delays Tj are introduced and systematically modified until the signature is discovered with reasonable magnitude in all (or a majority) of the channels.
  • the magnitude is measured by a mean (or RMS) across the atom coefficients in the collection of structure books.
  • Timing information is then extracted from the structure books for each atom and summed with the respective pre-delay value X 1 before being averaged to a final estimate of the arrival time in each channel. Location information is then estimated from the relative arrival times.
  • Fig.9B introduces pre-delays 99, 100, 101 before the SSA operation; the diagram is otherwise similar.
  • the matched structure books 98 will, however, differ since they interact with the pre delay times ⁇ j.
  • These independently variable time delays may be inserted to compensate for variations in arrival time at transducers so that the time-similarity window may be decreased in the GAD SSA to reduce false associations.
  • the delays may comprise actual time shifts introduced before the SSA or may be coded as time shifts that occur on the fly during the comparison operations of the SSA.
  • the time phase information 97 extracted from the structure books 96 may be used to guide the time delays ⁇ j (99, 100, 101) in order to "zero-in" on a proper window for each channel. (This idea is not illustrated in the figure.)
  • amplitude information may be extracted. This is not shown on the figure. Amplitude may be used to help model the source location and increase performance of the system through well-known methods. In addition, other parameters such as frequency and scale variations may inform the system about the environment. For example, different delays at different frequency ranges might indicate properties of materials through which the signal is traveling.
  • an additional novel method of extracting position information is available within the GAD framework. Rather than mathematically deducing location from time, amplitude, and other parameters, in certain embodiments, the system may be trained. By exposing signature extraction mechanisms (see Fig 3 or Fig 4, for example) to multiple presentations of a sound source in multiple locations, one may extract parametric mean representations of signature-location combinations. One may likewise present multiple signatures first at one location, than another, do utilize the system to discover automatically how sound characteristics change with the location of the source.
  • signature-location combinations are encoded just like signatures, hence the detectors / classifiers already described will return a map to both the signature and its location.
  • location is encoded separately, hence detection of signatures and localization will proceed independently (either sequentially or in parallel) using multiple detectors as described above.
  • Fig. 10 illustrates the training process for the two embodiments.
  • the process generalizes the basic training idea of sampling multiple sources to build a dictionary of signatures.
  • This novel idea also affords a deployed system the ability to learn its environment. That is, sensors may be positioned and exposed to sounds at known locations to calibrate the system and automatically compensate for environmental inhomogeneity.
  • the GAD engine makes available phase, time, and amplitude information for detected signatures. With proper dictionary choices, it also extracts frequency, phase, and other parameters. It will be clear to those skilled in the art that additional direction finding technique not enumerated here utilize such information, and thus may be facilitated by adapting methods similar to those disclosed.
  • the acoustic systems and methods may specifically be applied to a host of applications.
  • Some contemplated applications include: mineral discovery by monitoring acoustic returns from ground penetrating energy; detection and classification of personnel, vehicles, or other targets of strategic military or intelligence interest; detection and classification of personnel or equipment within buildings and structures; detection and classification of navel vessel, swimmers, divers, or marine animals in an underwater environment; detection and classification of human, animal, or mechanical sounds in remote wilderness; detection and classification of acoustic sounds in urban and other human occupied environments; detection and classification of aircraft and birds; detection and classification of weather phenomena; detection and classification of machine malfunction or impending malfunction; detection and classification of ultrasound returns in inspection of organisms or materials; detection and classification (diagnoses) of human health status and disease states; acoustic biometric; voice and/or speaker recognition.
  • This list is intended to be merely exemplary and should not in anyway be construed as exhaustive.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Computational Linguistics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
  • Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
  • Examining Or Testing Airtightness (AREA)

Abstract

Cette invention concerne un système et un procédé permettant de classifier et d'identifier des signaux acoustiques en fonction de la nature et de la position du signal d'origine. Le système et le procédé permettent de déterminer, à partir d'un ensemble arbitraire de signaux, une signature ou tout autre élément caractéristique et de différencier des signaux associés à plusieurs conditions au moyen de dictionnaires comprenant des atomes de signaux.
PCT/US2006/010892 2005-03-23 2006-03-23 Systeme et procede permettant d'extraire, de detecter, de differencier et de localiser une signature acoustique WO2006107624A2 (fr)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US66451805P 2005-03-23 2005-03-23
US60/664,518 2005-03-23
US11/387,034 2006-03-22
US11/387,034 US8271200B2 (en) 2003-12-31 2006-03-22 System and method for acoustic signature extraction, detection, discrimination, and localization

Publications (2)

Publication Number Publication Date
WO2006107624A2 true WO2006107624A2 (fr) 2006-10-12
WO2006107624A3 WO2006107624A3 (fr) 2008-11-13

Family

ID=37073940

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2006/010892 WO2006107624A2 (fr) 2005-03-23 2006-03-23 Systeme et procede permettant d'extraire, de detecter, de differencier et de localiser une signature acoustique

Country Status (2)

Country Link
US (2) US8271200B2 (fr)
WO (1) WO2006107624A2 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020220091A1 (fr) * 2019-05-02 2020-11-05 ResMed Pty Ltd Identification de composant acoustique pour systèmes de thérapie respiratoire

Families Citing this family (38)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7872574B2 (en) * 2006-02-01 2011-01-18 Innovation Specialists, Llc Sensory enhancement systems and methods in personal electronic devices
DE602006002132D1 (de) * 2006-12-14 2008-09-18 Harman Becker Automotive Sys beitung
US8325562B2 (en) 2007-02-09 2012-12-04 Shotspotter, Inc. Acoustic survey methods in weapons location systems
GB0712936D0 (en) * 2007-07-05 2007-08-15 Airbus Uk Ltd A Method, apparatus or software for determining the location of an acoustic emission emitted in a structure
JPWO2009051132A1 (ja) * 2007-10-19 2011-03-03 日本電気株式会社 信号処理システムと、その装置、方法及びそのプログラム
US20100082180A1 (en) * 2008-10-01 2010-04-01 Honeywell International Inc. Errant vehicle countermeasures
US8983677B2 (en) * 2008-10-01 2015-03-17 Honeywell International Inc. Acoustic fingerprinting of mechanical devices
US8331195B1 (en) * 2008-10-20 2012-12-11 The United States Of America As Represented By The Secretary Of The Army Computer implemented sensor data analysis
JP5101465B2 (ja) 2008-11-25 2012-12-19 三菱重工業株式会社 設備の不具合管理方法
JP5622744B2 (ja) * 2009-11-06 2014-11-12 株式会社東芝 音声認識装置
US8380389B2 (en) * 2009-11-30 2013-02-19 Honeywell International Inc. Health monitoring systems and methods with vehicle identification
US8805083B1 (en) * 2010-03-21 2014-08-12 Jeffrey M. Sieracki System and method for discriminating constituents of image by complex spectral signature extraction
US8478319B2 (en) 2010-05-12 2013-07-02 Information System Technologies, Inc. Feature extraction and data compression system and method for distributed sensor networks
US8784311B2 (en) * 2010-10-05 2014-07-22 University Of Florida Research Foundation, Incorporated Systems and methods of screening for medical states using speech and other vocal behaviors
US9558762B1 (en) * 2011-07-03 2017-01-31 Reality Analytics, Inc. System and method for distinguishing source from unconstrained acoustic signals emitted thereby in context agnostic manner
US9886945B1 (en) * 2011-07-03 2018-02-06 Reality Analytics, Inc. System and method for taxonomically distinguishing sample data captured from biota sources
US9691395B1 (en) * 2011-12-31 2017-06-27 Reality Analytics, Inc. System and method for taxonomically distinguishing unconstrained signal data segments
US9939540B2 (en) * 2011-08-19 2018-04-10 Cornell University System and methods for remote monitoring
AT510674B1 (de) * 2011-10-27 2014-05-15 Avl List Gmbh Verfahren und eine vorrichtung zum parametrieren eines sensors
US9070012B1 (en) 2012-01-11 2015-06-30 Sr2 Group, Llc System and method for uncued discrimination of bated features in image
US9183844B2 (en) * 2012-05-22 2015-11-10 Harris Corporation Near-field noise cancellation
US8959022B2 (en) 2012-07-03 2015-02-17 Motorola Solutions, Inc. System for media correlation based on latent evidences of audio
JP6557219B2 (ja) * 2013-09-27 2019-08-07 コーニンクレッカ フィリップス エヌ ヴェＫｏｎｉｎｋｌｉｊｋｅ Ｐｈｉｌｉｐｓ Ｎ．Ｖ． 生理信号を処理する処理装置、処理方法及びシステム
US9567847B2 (en) * 2013-10-02 2017-02-14 Hannes Wedemeyer Method of data volume reduction for transmission on a limited communications channel
US9613626B2 (en) * 2015-02-06 2017-04-04 Fortemedia, Inc. Audio device for recognizing key phrases and method thereof
US9805739B2 (en) * 2015-05-15 2017-10-31 Google Inc. Sound event detection
US10032464B2 (en) * 2015-11-24 2018-07-24 Droneshield, Llc Drone detection and classification with compensation for background clutter sources
US9959885B2 (en) * 2016-03-08 2018-05-01 Samsung Eletrônica Da Amazônia Ltda Method for user context recognition using sound signatures
US10141005B2 (en) 2016-06-10 2018-11-27 Apple Inc. Noise detection and removal systems, and related methods
US11074828B2 (en) * 2016-12-13 2021-07-27 The Boeing Company Sonar target simulator
US20180375444A1 (en) * 2017-06-23 2018-12-27 Johnson Controls Technology Company Building system with vibration based occupancy sensors
DE102018208118A1 (de) * 2018-05-23 2019-11-28 Robert Bosch Gmbh Verfahren und Vorrichtung zum Authentifizieren einer über einen Bus übertragenen Nachricht
EP3722998A1 (fr) * 2019-04-11 2020-10-14 Teraki GmbH Analyse de données sur des signaux prétraités
US11231815B2 (en) 2019-06-28 2022-01-25 Qeexo, Co. Detecting object proximity using touch sensitive surface sensing and ultrasonic sensing
US11592423B2 (en) * 2020-01-29 2023-02-28 Qeexo, Co. Adaptive ultrasonic sensing techniques and systems to mitigate interference
US12211516B2 (en) 2020-03-10 2025-01-28 Eaton Intelligent Power Limited Noise event detection and characterization
WO2022056104A1 (fr) * 2020-09-09 2022-03-17 Kai Systems, Inc. Système autonome de surveillance de l'intégrité structurale de pieux marins
TWI859581B (zh) * 2022-09-15 2024-10-21 中國鋼鐵股份有限公司 具有週期性動作之設備的監控方法

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE69202699T2 (de) * 1991-02-07 1996-01-18 Thomson Multimedia Sa Verfahren, Identifizierungs- und Verifiziereinrichtung zur Identifizierung und/oder Ausführung digitaler Unterschriften.
US5699121A (en) * 1995-09-21 1997-12-16 Regents Of The University Of California Method and apparatus for compression of low bit rate video signals
US5764921A (en) * 1995-10-26 1998-06-09 Motorola Method, device and microprocessor for selectively compressing video frames of a motion compensated prediction-based video codec
US6587507B1 (en) * 1997-06-02 2003-07-01 Picsurf, Inc. System and method for encoding video data using computationally efficient adaptive spline wavelets
US6016546A (en) * 1997-07-10 2000-01-18 International Business Machines Corporation Efficient detection of computer viruses and other data traits
US6148106A (en) * 1998-06-30 2000-11-14 The United States Of America As Represented By The Secretary Of The Navy Classification of images using a dictionary of compressed time-frequency atoms
US7020701B1 (en) * 1999-10-06 2006-03-28 Sensoria Corporation Method for collecting and processing data using internetworked wireless integrated network sensors (WINS)
KR20010105362A (ko) * 1999-12-28 2001-11-28 요트.게.아. 롤페즈 정합 추적 알고리즘에 기초한 비디오 인코딩 방법
WO2001049036A1 (fr) * 1999-12-28 2001-07-05 Koninklijke Philips Electronics N.V. Procede de codage video a rapport signal-bruit echelonnable et procede de decodage correspondant
EP1338001B1 (fr) * 2000-11-03 2007-02-21 Koninklijke Philips Electronics N.V. Codage de signaux audio
JP2002315004A (ja) * 2001-04-09 2002-10-25 Ntt Docomo Inc 画像符号化方法及び装置、画像復号方法及び装置、並びに画像処理システム
US7003039B2 (en) * 2001-07-18 2006-02-21 Avideh Zakhor Dictionary generation method for video and image compression
US6628300B2 (en) * 2001-11-30 2003-09-30 International Business Machines Corporation Transcoding proxy and method for transcoding encoded streams
US7006567B2 (en) * 2001-11-30 2006-02-28 International Business Machines Corporation System and method for encoding three-dimensional signals using a matching pursuit algorithm
US6944222B2 (en) * 2002-03-04 2005-09-13 Koninklijke Philips Electronics N.V. Efficiency FGST framework employing higher quality reference frames
JP2005051734A (ja) * 2003-07-15 2005-02-24 Hitachi Ltd 電子文書の真正性保証方法および電子文書の公開システム

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
BULTAN A.: 'A Four-Parameter Atomic Decomposition of Chirplets' IEEE TRANSACTIONS ON SIGNAL PROCESSING vol. 47, March 1999, pages 731 - 745 *
MALLET S.G.: 'Matching pursuits with time-frequency dictionaries' IEEE TRANSACTIONS ON SIGNAL PROCESSING vol. 41, December 1999, pages 3397 - 3415 *
SANDLER B.: 'Optimal and wavelet-based shock wave decomposition and estimation' J. ACCOUSTIC SOC. AM. vol. 104, August 1998, pages 955 - 963 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020220091A1 (fr) * 2019-05-02 2020-11-05 ResMed Pty Ltd Identification de composant acoustique pour systèmes de thérapie respiratoire

Also Published As

Publication number Publication date
US8271200B2 (en) 2012-09-18
WO2006107624A3 (fr) 2008-11-13
US20130073263A1 (en) 2013-03-21
US20060241916A1 (en) 2006-10-26

Similar Documents

Publication Publication Date Title
US8271200B2 (en) System and method for acoustic signature extraction, detection, discrimination, and localization
Tak et al. End-to-end anti-spoofing with rawnet2
US9666183B2 (en) Deep neural net based filter prediction for audio event classification and extraction
Stöter et al. CountNet: Estimating the number of concurrent speakers using supervised learning
Skowronski et al. Acoustic detection and classification of microchiroptera using machine learning: lessons learned from automatic speech recognition
Socoró et al. Development of an Anomalous Noise Event Detection Algorithm for dynamic road traffic noise mapping
Vuong et al. Learnable spectro-temporal receptive fields for robust voice type discrimination
Mouy et al. Automatic recognition of fin and blue whale calls for real-time monitoring in the St. Lawrence
KR101082839B1 (ko) 다채널 잡음처리 장치 및 방법
Gul et al. Clustering of spatial cues by semantic segmentation for anechoic binaural source separation
Keronen et al. Mask estimation and imputation methods for missing data speech recognition in a multisource reverberant environment
Wang et al. Multi-speaker speech separation under reverberation conditions using Conv-Tasnet
Fernandes et al. Underwater target classification with optimized feature selection based on genetic algorithms
Kamble et al. Teager energy subband filtered features for near and far-field automatic speech recognition
Hajihashemi et al. Novel time-frequency based scheme for detecting sound events from sound background in audio segments
Evans Automated vehicle detection and classification using acoustic and seismic signals
Ahmed et al. Detecting replay attack on voice-controlled systems using small neural networks
Patil et al. Significance of cmvn for replay spoof detection
Vickers et al. A comparison of machine learning methods for detecting right whales from autonomous surface vehicles
JP7000963B2 (ja) ソーナー装置、音響信号判別法、及びプログラム
Silva et al. A wavelet transform-based feature extraction pipeline for elephant rumble detection
Neri et al. Multi-channel Replay Speech Detection using an Adaptive Learnable Beamformer
Guzewich et al. Cross-Corpora Convolutional Deep Neural Network Dereverberation Preprocessing for Speaker Verification and Speech Enhancement.
Kühne et al. Time-frequency masking: Linking blind source separation and robust speech recognition
Hoyt et al. Detection of human speech using hybrid recognition models

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application
DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
NENP Non-entry into the national phase

Ref country code: DE

NENP Non-entry into the national phase

Ref country code: RU

122 Ep: pct application non-entry in european phase

Ref document number: 06758194

Country of ref document: EP

Kind code of ref document: A2

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载